Introduction

Critical Periods

Although often associated with childhood, learning is an essential part of life that allows people to adapt to their ever-changing environment even after childhood. This capacity to learn new skills and adapt knowledge is due to the plasticity of the brain. Plasticity refers to the brain’s ability to change the connectivity of its neurons as a result of outside stimuli or fluctuations (Lillard, & Erisir, 2011). During postnatal development there are periods of great plasticity called critical or sensitive periods, during which the high sensitivity of neurons leads to drastic changes in the brain (Lillard & Erisir, 2011; Hübener & Bonhoeffer, 2014; Hensch, 2004). The plastic nature of the brain during the first few years of life allows it to make new connections and disconnect ones deemed unfit, through both active and passive exposure (Lillard & Erisir, 2011; Hübener & Bonhoeffer, 2014). Critical periods are not universal to the brain, as they vary between modalities which develop at different rates with different cascaded pathways (Hensch, 2004). Previously, it was believed that once critical periods passed, it was not possible to continue development, since brain plasticity decreases drastically with age (Hübener & Bonhoeffer, 2014; Hensch, 2004). This view of the critical period has shifted towards one of a sensitive periods, as research has shown that critical period sensitivity can be instigated beyond the ‘critical’ window under the right conditions, such as intense stimuli or chemical changes in the brain (Hensch, 2004). esearch continuously focuses on the reopening of critical periods or creation of sensitive periods in an effort to improve adult learning.

Absolute Pitch Critical Periods

Absolute pitch has a critical period that is often targeted in research, because its critical period is unusually long in humans, and therefore more easily studied. Absolute pitch (AP) is defined as the ability to accurately identify the pitch of a musical tone or to produce any of the 12 musical tones in the Western Musical System at a given pitch without the use of an external reference (Takeuchi & Hulse, 1993). While AP is rare, more so in western cultures than eastern, many individuals have and use relative pitch (RP) (Profita, Bidder, Optiz, & Reynolds, 1988). RP is the ability to identify tones by processing the melodic and harmonic relations between pitches. RP and AP vary in that RP has been identified as an ability that can be continuously improved, while AP is not (Gervain et al., 2013).

The existence of a critical period for absolute pitch is supported by evidence from previous studies which identified the likelihood of AP acquisition based on the age of onset of musical training (Deutsch et al., 2009); in other words, the presence of AP ability is dependent on the age at which an individual starts their musical training. The AP skill was identified mostly in individuals who began their musical training between the ages of 4-6 (Levitin & Zatorre, 2003), and the number of individuals acquiring the skill decreased with the age of onset of musical training. To date, there has been no record of an adult who has acquired the ability (Gervain et al., 2013). Therefore, exposure to musical training before the closure of the critical period is crucial for the acquisition of absolute pitch.

Amongst individuals who began their musical training before the age of 6, there is a mix of those who possess AP and those who do not. There is also a spread in terms of ability levels; not all AP possessors are able to identify the pitch class of a sound produced by instruments other than the one(s) on which they were trained (Bermudez, 2008). From this, it can be said that absolute pitch is a continuum. The differences in ability is believed to exist in part, due to the differences in the type of training (Gervain et al., 2013). Overall, AP possessors have been found to be most successful in identifying pitch class between sounds produced by the instrument that they are most familiar with, usually an acoustic piano (Bermudez, 2008). In terms of the methods of training for absolute pitch acquisition, individuals need to learn and practice the association of pitches to their respective labels during the early stages of their training. The exposure to different pitches and the knowledge of their labels for purpose of learning how to play an instrument or read music is not enough to successfully acquire perfect pitch (Lockhead & Byrd, 1998). Without the specific training of pitch class and label association, the perceptual system of auditory input is reorganized; the neural architecture shifts weight from absolute to relative pitch (Takeuchi & Hulse, 1993). In other words, without the presence of training to learn the specific AP cues during the critical period, individuals are most likely to develop highly trained relative pitch, which is found in many non-AP musicians.

In addition to differences in information processing, differences in brain structures has also been identified. One difference is in corpus callosum size between musicians and non-musicians, the structure being bigger in the first group, as well as hyperconnectivity in the temporal cortex (Loui et al., 2011; Schlaug, Jäncke, Huang, & Staiger, 1995) in AP possessors. A specific structure in the temporal cortex that has been identified to be different in AP and AP non-possessors is the planum temporale, which is associated with learning conditional associations (Zatorre, 2003). This is significant since has been found that the skill involves the association of labels to pitch classes in long-term memory (Zatorre, 2003).While the left planum temporale is larger than the right one in AP possessors, no significant difference between the left and right counterparts in AP non-possessors. Although it has been found that AP possessors are able to easily identify and produce any one of the 12 pitches, they are not always correct; they do, however, complete the task with much greater success than individuals without the ability. Absolute pitch (AP) possessors generally succeed in identification or production of tones 70% - 99% of the time, compared to the 10% - 40% success rate of AP non-possessors (Takeuchi & Hulse, 1993). Although AP possessors correctly identify chroma, or pitch class, a common mistake made is in the identification of pitch height, or octave. From this, it is inferred that pitch class and pitch height identification involve different processes (Levitin & Rogers, 2005). For the purpose of this study, we will focus on the accuracy of pitch class identification rather than pitch height.

Use of Drugs for Critical Period Reopening

Currently there are several discussed methods for attempting to reopen critical periods, each corresponds to different theories regarding the causes of critical period closure. In this proposed research, we are focusing on the theory that explains closing of critical periods through neurotransmitter inhibition(Hensch, 2004; Hensch & Bilimoria, 2012; Gervain et al. 2013). There are many scholars who theorize and look at drugs as a tool to manipulate and regulate human critical periods. Hensch (2004), specifically gives great insight of the various neurotransmitters and the systems they affect; Bellone & Lüscher (2012), outlines the potential of cocaine as an increaser of neural plasticity; Hensch, Fagiolini, Mataga, Stryker, Baekkeskov and Kash (1998) use benzodiazepine to trigger plasticity in ocular dominance of mice. Hensch (2004) explains that there is an important role in both the interconnectedness of neuronal circuits and also the inhibition of central nervous system (or more specifically, its neurotransmitters). An example of such is found in how the use of an NMDA-glutamate blocker resulted in a sharp 2 day critical period in lab rats (Hensch 2004). It should be noted that this is not an exhaustive list. The proposed research will be expanding a 2013 study by Gervain et al. This study found that using Valproate (VPA), an HDAC inhibitor used in mood stabilization and epilepsy treatment, they may have re-opened the critical period for absolute pitch. Gervain et al.’s experiment results found that those who took VPA scored significantly higher in identifying note pitches than those who did not take VPA (2013). Gervain et al.’s study is the impetus for the proposed study and there will be further steps taken to analyze VPA’s potential effects on critical periods.

Neural Networks as Information Processing Tools

Neural networks are a prominent topic in modern computing, and for good reason: the underlying concepts behind them are relatively simple and straightforward, yet over the last decade, they have repeatedly set the standard with state of the art results in a wide variety of tasks, from speech recognition ("Historic Achievement: Microsoft researchers reach human parity in conversational speech recognition - The AI Blog", 2017) to image classification (Krizhevsky et al. (2012)). It should be noted that image classification is the task they will be used for in this study. The type of neural network that will be used in this study will be a novel new architecture, the capsule network or CapsNet (from the very recently published paper “Dynamic Routing Between Capsules” by Sabour et al. (2017)), which makes several interesting and promising departures from existing neural network designs.

Research Aim

Our goal is to look at AP critical periods and research done into reopening of this critical period. Specifically the proposed research looks at the use of the drug Valproate (VPA), and how similar the sensitivity created by VPA critical period is to that of the normal AP critical period. This implies analyzing the similarity in stimuli perception and processing by people who acquired AP through the normal critical period and those who acquired it through VPA.

Hypothesis

In terms of brain activity data, it is hypothesized that our findings will be similar to those found in previous studies. The brain activity data collected during AP testing after VPA trials is expected to be analogous to that of normal AP acquisition. The differences in patterns between AP and non-AP possessors identified after data processing through the CapsNet should be noticeably different.

Methodology

Summary

The experiment will gather fMRI data of male subjects between the ages of 18-30 (outside of AP critical period) who underwent AP training with and without VPA and compare it with fMRI data of people who acquired AP during its theorized critical period. The comparison will be done by a neural network trained on fMRI data of people with AP and RP. Therefore, the experiment is divided into two main sequential stages, with a preliminary data gathering and recruitment stage (Stage Zero). Stage One consists of gathering data for the training of the neural network. Stage Two consists of administering VPA or a placebo to the subjects, as per Gervain et al., then gathering fMRI on these subjects, and finally running the data through the neural network.

Population & Sample

Our target population is adult males, ages 18 to 30. The AP musicians must have classical piano training to level 10 of the Royal Conservatory of Music. The AP musicians should also have started their training from ages 3 to 5 and still be active with music at the time of recruitment. Classical training has been chosen as this school teaches the use of sheet music, which does not necessarily lend itself to relative pitch enhancement. This is contrasted by contemporary schools that teach students to play by ear (Elite Music, 2013) and may lead to highly skilled relative pitch musicians. The non-AP participants will need to have no musical training.

The sample size will be at least 200 subjects, half of which are AP possessors (those with AP) and the other half AP non-possessors (those without AP). This sample size is significantly larger than both Gervain et al. (2013) and slightly larger than Deutsch et al. (2006), due to the neural network requiring a large data pool on which it can be trained. Moreover, a larger sample size gives greater accuracy, as long as it is representative of the population. Recruitment will be done by stratified sampling at the UBC and SFU music programs for AP possessors, and through an advertisement for non-AP recruitment. While the latter will be done through public advertisement, there will be requirements listed and these will control for the multitude of variables listed above.

Test Groups

The participants will be split into 2 main groups – AP possessors and AP non-possessors. The latter group (non-AP) will be broken down into 2 even subgroups when the study moves into Stage 2; they will be split into VPA-treated (AP non-possessors who receive VPA treatment) and Placebo (AP non-possessors who do not receive VPA treatment).

Variables

Independent

The independent variable is the treatment each test group will undergo, specifically VPA and non-VPA.

Dependent

The dependent variable is brain activity and the pattern correlation found by the neural network to brain activity of normal AP acquisition.

Control

Control variables include age (18-30), gender/sex (male), languages spoken, and handedness (right handedness). Handedness will be tested using the Edinburgh Handedness Inventory (Oldfield, 1971). The use of monolingual English-only speakers for non-AP participants is very important. There would be an extremely long recruitment process for that group if tonal language speakers were not accounted for. Another requirement will be that subjects are physically healthy, since any health issues may interfere with the administration of Valproate. The use of the Valproate is also the reason for controlling gender, due to issues found with pregnancy and VPA (Drugs.com, 2017; Gervain et al. 2013).

Compensation

Participants will be compensated for their time. They will be paid 10% higher than the British Columbia minimum wage (minimum wage is $11.35/hour as of November 2017 thus they would be paid $12.49/hour), which we believe is fair for their effort. At a minimum, participants will get at least 1 hour’s worth of pay when they visit (this is mainly aimed at the non-AP group, as they will need to show up multiple times). Participants will start receiving payment during their first session in the study, and will receive payment per session.

Procedure/Methods

Stage Zero

Preliminary AP Test

There will be a preliminary AP test that every potential participant will have to take. This test will judge AP ability (or a lack thereof). For non-AP people, it will serve to test their level of RP. The preliminary AP test to be used is a modified version of previously used AP tests. (Deutsch et al. 2006; Gervain et al., 2013). This test will serve as a check of “raw” ability, as such there will be no pre-training for either group. The physical design is as follows: the subject will be given studio quality headphones that are adjustable, to accommodate head size. A monitor will display what question the subject is on; and a button response system made to be a physical replica of one contained in the MRI. As with Deutsch et al. the test will use 36 (12 tones and 3 rounds) western tones in intervals of octave. The tones will be split into 4 even blocks for ease of participants. The tones will go an octave up each time, and they will be played in intervals of 4.25 seconds to give participants a chance to quickly answer, and 20 seconds between each block. Each tone will have an attack and decay (fade in/fade out) of 30ms and randomly have distractor tones (notes not of the western scale) dispersed in-between (Schulze, 2009). This test will be computerized, since the paper response design is cumbersome and problematic as it allows for more human errors, such as indiscernible writing, losing responses. As per Gervain et al. (2013), the freely available software “Psyscope X” will be used to design a test in which participants are asked to listen, and then to respond to a sound stimuli. Our sound production will be done through an acoustic piano and using a directional microphone, each tone will be recorded to an uncompressed format (specifically, to raw audio files) and backed up onto a 4.7GB DVD. The test will not use pure sine and cosine waves as this has been found to skew data in odd ways – a study found that AP and non-AP were both found to be dubiously accurate in their answers when the researchers used sine and cosine waves (Hirose et al., 2002). Despite the lengthy description of how the test will be done, the actual test itself should be quite short in comparison to the fMRI AP tests that are to follow.

Stage One

fMRI AP Test #1

The fMRI design will be based on the already established method/setting done by Schwenzer and Mathiak (2011). Like them, we will use a single-shot triple-echo EPI pulse sequence, a repetition time of 6 seconds, an echo times of 17, 43 and 68 milliseconds, an acquisition time of 2.8 seconds, 30 slices acquired per cycle, voxel size of 3.6*3.6*4mm and a matrix size of 64x56 will be used for the fMRI settings (Schwenzer and Mathiak, 2011). To further break the timing down, it seems that Schwenzer and Mathiak set a temporal gap; (TR[6])-(TR[6]/N[30]) = 5.8 where TR is repeat time in seconds, N is slices. Schwenzer and Mathiak find success in identifying BOLD (blood oxygenation level-dependent) activity during tone identification using these settings and thus we believe they are adequate for the proposed research as it too will require BOLD activity imaging. The proposed research will diverge in various places, both from necessity and choice. Differences will be found in the MRI used; instead of the Siemens imaging machine Schwenzer and Mathiak (2011) used, UBC’s Philips Achieva 3.0 Tesla whole body scanner will be used (coincidentally, the Philips machine seems easier to understand, and we will have staff on hand). Furthermore, one of the most drastic differences is that rather than use SPM2 to analyze the data (as Schwenzer and Mathiak did), the obtained data will be sent through a neural network for analysis instead. There will be AP testing during the fMRI data collection. This testing will follow the same procedure as the preliminary test, albeit slightly modified as the screen will be a projection into the MRI.

Neural Network Training

The data collected in Stage One will be divided into being from AP or non-AP subjects and labelled accordingly for processing in the CapsNet manually. The hyperparameters of the CapsNet will be set according to the Bayesian optimization guidelines and techniques in the paper “Practical Bayesian Optimization of Machine Learning Algorithms” by Snoek et al. (2012). If no academic computing resources are available, training will be conducted on a commercial cloud computing service such as Amazon Web Services EC2.

Stage Two

Valproate Administration and Training

Due to the large size of the VPA treatment group, there will need be a staggered beginning for treatment. This will be done by breaking the VPA treatment group down further into groups of 10 (for a total of 5 groups), internally labelled from 1-5, with each group having a different start date. The groups will have separate fMRI AP Test #2 dates as well. This staggering will be done to ensure each participant has had an overall similar amount of time for VPA to take effect. In other words, by the date of their individual Test #2s, everyone will have 15 days of VPA treatment. To keep consistency and the double-blind aspect intact, we will do the same for the placebo group. We will be following the same dosage amounts and schedule as laid out by Gervain et al. (2013). That is, the participants in stage 2 will undergo 15 days of taking a pill (whether it be VPA or a placebo). Days 1 to 3 will be 250mg at strictly morning and night (for a total of 500mg); this will turn into 1000mg for days 4 to 14 (250mg in the morning, 250mg in the afternoon and 500mg at night); and on the 15th day of trials – also the day that both the VPA treatment and placebo groups take Test #2 – they will take 250mg in the morning. This design is based off of previous study (Bell 2005 in Gervain et al., 2013) that complies with Health Canada’s safety standards. During and after the VPA-treatment procedure, we will monitor the participants’ health (both mental and physical). We will also look for any behavioral changes. If anything adverse occurs, we will interfere and cancel their participation and refer them to a hospital/medical professional.

Training for AP

Non-AP participants will be required to undergo training in order for them to learn musical pitch. The training design will follow that of Gervain et al.’s (2013) in the way they associated different musical tones with names; this is a novel approach that should make the test more palatable for the non-AP people. However, will diverge in a few specific ways as we will need to consider the ease of us for two distinct groups (AP and non-AP) rather than just one. Rather than associate the pitches with random western names, we will assign names that coincide with the tone letters (e.g. C will be Cynthia, D will be Daniel). The names chosen will be different in training and the actual AP test.

fMRI AP Test #2

Whereas participants were instructed to do an online training session for 7 days before the final AP test in Gervain et al. (2013), the proposed study will already have trained (and refreshed) the participants twice before this. We will test AP ability in the same way we did in step 2.

Processing of fMRI Data for AP Test

The data collected in Stage Two is run through the trained CapsNet from the end of Stage One. During each computation, the output of each capsule in the CapsNet is recorded. The resulting classification and classification probability are also recorded. After all of the results have been computed, the vectors recorded from the capsules for AP and non-AP classifications will be compared using common statistical analysis tools such as Excel, Python, R, etc.

Discussion

Ethics

General

Before proceeding with the experiment, participants will be given a written consent form and will also be given an informed through a verbal explanation of what the study entails. The consent form will conform to instructional policies/ethics guidelines as well as national health guidelines (we are administering a drug, after all). We will honor participants’ anonymity and keep their identities confidential/anonymous. We will do this “naturally” through the research design by not reporting any single individual’s results, but rather an amalgamated result. If, for some reason, the study were to be sponsored by a biased third party, we would not commence. We will require both AP and non-AP participants to disclose if they have any instances of foreign/metallic objects in their body – if these objects are non-ferrous they will be accepted, otherwise they will not be accepted as the MRI section may pose a risk; participants will be referred to UBC’s checklist on how to prepare for MRIs. We will require the non-AP participants to further get a thorough health inspection checking their eligibility for VPA – if they do not receive a clean bill of health, they will not be considered eligible for the research at all; our Stage 2 group distribution relies random selection, and thus on all of the non-AP members to be healthy enough for VPA.

Valproate

VPA is administered to treat bipolar disorder, epilepsy, schizophrenia, whilst also being used as a preventative tool for seizures and migraines (Drugs.com, 2017). With the use of VPA we will need to be aware of all its usage issues, specifically the fact that we cannot include women in the study as one of its risks are birth defects, as well as the side effects (albeit rare) of hepatotoxicity and suicidal thoughts (Drugs.com, 2017), amongst other issues. We will be as vigilant as possible, and monitoring the overall physical and mental health of participants who undergo the VPA treatment stage. As with any study, participants can quit at will – we will not coerce the involvement of anyone – however, we will request that those who undergo VPA treatment get follow ups/check ups to measure their health. There should be no issue with participants’ discontinuation of VPA treatment (this is generally only an issue with epileptics) (Drugs.com, 2017), but we would like to be thorough – health is not something to be taken lightly. Similarly, if any participant shows or feels adverse effects, we will discontinue treatment immediately and if it is within our immediate vicinity, we will call an ambulance to take them to a hospital.

Data Collection

Before deciding to go with fMRI, multiple other methods were considered. The front runners amongst those methods were EEG and ECoG. We looked at a combination of fMRI with EEG or ECoG, but fMRI-ECoG was discarded early as it was deemed to be much too invasive. Ihalainen et al. (2015) looks at the combination of fMRI/EEG and finds that the procedure will lead to artifacts on the fMRI’s produced images (albeit superficial), though (somewhat predictably) there is great potential for a much richer dataset. However, Schwenzer and Mathiak’s 2011 study shows that, with the correct settings, fMRI is a more than adequate method for our purpose and we have decided (while keeping our original intent in mind) that including the other methods may lead to (a mild level of) scope creep. The fMRI-only method has its advantages in how it allows the observation of brain activity centers – far more valuable than the quickness advantage of EEG. fMRI nowhere near as invasive as ECoG, and the data is also easier to understand when compared to both ECoG and EEG. fMRI has good spatial resolution (Pylkkanen, 2016). Once it is done going through the MRI output, our neural network will further improve the visual aspect of the fMRI technique and the neural network will ideally provide great visuals with more easily digestible data. However, fMRI is disadvantageous due to its low temporal resolution in comparison to the millisecond temporal resolution of EEG/ECoG, however fMRI is fast enough for us to tell the difference between our various trials (Pylkkanen, 2016). fMRI also cannot discriminate between normal (baseline) brain activity and those exclusively responsible for AP activities; rather, it only indicates areas of increased activity. Furthermore fMRI could be problematic, on a practical level, for participants who are claustrophobic. Another issue with fMRI is found with requiring participants to lay still for prolonged periods of time; this creates issues as participants will undoubtedly move and create variation in voxel data (Notter, 2016; Preston, 2006). This can be mostly solved through realignment of axes – 3 dimensional axes and 3 rotational axes – as well as artifact detection (i.e. the removal of certain images) (Notter 2016). Due to the movement of participants, the data will be subject to some unavoidable level of deterioration that we can only try to account for.

The MRI in question is UBC’s Philips Achieva 3.0 Tesla. It is well equipped for adequate for our visual, audio and response needs. The Achieva is equipped with the following: a projection system that to provide the visual aspect of our AP test; an audio presentation system for audio playback; a touch/button response system, and the latter two are absolutely imperative for our experiment design. The fact that the Philips machine has a stronger 3.0t magnet compared to the 1.5t Siemens device has been considered, and some accomodating may be necessary for the differences in image artifact that might occur. The convenience of using the UBC scanner outweighs the minor artifacts that may occur in a higher powered machine. The use of a different device is not of great concern, as the experiment is not meant to replicate neither Schwenzer and Mathiak’s (2011) nor Schulze et al.’s (2009) results.

Data Analysis

Of course, neural networks are not the only viable machine learning technique for image classification. Methods such as support vector machines, random forests, and recursive cortical networks have also produced respectable ("MNIST handwritten digit database, Yann LeCun, Corinna Cortes and Chris Burges", 2017) results with comparable amounts of effort needed for setup. However, of the methods popular today, only neural networks and recursive cortical networks are modelled to some extent after the human brain. Recursive cortical networks are as of yet mostly unproven beyond a recent impressive CAPTCHA breaking performance ("This AI Technique Was Kept Quiet so Spammers Wouldn’t Misuse It", 2017). Neural networks are easily the most accessible method, as a wealth of resources are available chronicling and explaining cutting-edge discoveries and methods practically as soon as they appear. Finally, CapsNets promise superior performance and ease of use to current state of the art neural network models. Previous approaches to processing images with neural networks typically involved convolutional neural networks (CNNs), introduced by LeCun et al. (1998). They work by applying convolutions to an input image by sliding matrices (aka “kernels” or “filters”) across the image, applying a nonlinear activation function to each pixel in the convolved images (aka “feature maps”), “pooling” each feature map by shrinking it in size, then sending all of the feature maps to a fully connected (every neuron is connected to every neuron in front of it) layer of neurons that use the feature maps to compute a classification for the input image. Although CNNs are a proven and extremely popular architecture, they are ineffective at spatial context for the features they detect. That is, they are inadequate in discerning the correct size or orientation of a feature, or recognizing it from different perspectives, which means more data is required to compensate. Rather than CNNs, we will use the aforementioned capsule networks. The main innovation of capsule networks is that they preserve the spatial relationship between objects features in 4D “pose” matrices which describe the rotations and translations (i.e. the poses) between them. This is achieved by working with “capsules” of neurons which work with vectors rather than the scalar values used by individual neurons. The length of the vectors represent feature detection probabilities, and the direction they point encodes state information of the features detected. For fMRI data, where blood flow activity could look similar across different parts of the brain, it is practically a given that we will wish to take into account the spatial context of the features we detect. With that in mind, CapsNets appear to be one of the best, if not the best method of doing so, as they take a method that already defines the state of the art and improve on it in almost every way.

Limitations

There are limitations to the conclusions that can be drawn from this experiment. as a result of the experiment design and general biological variation between individuals. Firstly, the conclusions reached by such an experiment cannot be used to generalize to entire populations. This is in large part a result of the sample drawn; as it does not contain women nor adults over the age of 30. Generalizability is further limited by cultural aspects of the sample population, such as language spoken, since tonal languages seem to have a different effect on AP acquisition (Hou, Chen, Song, Sun & Beauchaine, 2016). Conclusions are also restricted to the AP critical period and cannot be generalized to that of other systems since critical periods vary between systems in timing, sequences, and chemical components. Therefore, while VPA may reactivate AP critical periods, it may not work for systems that have different neuroplasticity inhibitory mechanisms (Hensch, 2004). Moreover, it is not clear, whether all people who have AP learned it during the theorized critical period or whether it is an innate ability. If it is the case that AP possession has an innate component, it could affect the results, since some test subjects might simply be unable to learn it in or outside of the critical period. Furthermore if AP possession may be both an innate or learned could be problematic because not discriminating between the two when training the neural network may skew the results. Finally the technology for analyzing brain activity is still very limited. In the case of fMRI, there is the issue of limited temporal resolution mentioned earlier. While, the fMRI technique may be useful to determine brain areas activated, its limited low temporal resolution renders it insufficient for studying the dynamics of brain functions (Lewis, Bosman, & Fries, 2015). Special high temporal resolution fMRI techniques are needed for greater sensitivity of functional activation, as well as a greater ability to filter out noise (Petrov, Herbst & Stenger, 2017). The machine used in the experiment may not have this ability. In light of recent studies that highlight the role temporal multiplexing as a possible mechanism encoding and processing for sensory information, temporal resolution poses more limitations to the conclusions based on fMRI data (Panzeri, Brunel, Logothetis, & Kayser, 2010).

Further Research

Although the general areas of brain activation during chroma identification as well as differences in physical brain structures between AP possessors and non-possessors have been identified, the specific neural architecture and the actual processes involved in pitch identification in AP possessors is still being mapped and studied. This knowledge gap has prevented the development of methods for AP acquisition in individuals who have passed the AP critical period. In order to develop a method to successfully teach a non-AP possessor how to acquire the ability, we must first understand the differences between the two groups in as much detail as possible, which is the goal of our research proposal. By conducting this study, we hope to help close the knowledge gap in that area; the neural network database can act as an effective testing method to help further research on the reopening critical period. After the database of brain activity has been compiled, and patterns identified, it will be easier to conduct further research on learning methods. Although the Valproate drugs has been found to be effective, the side effects of the drug have been found to be very dangerous, making it an unethical and unrealistic pathway of AP acquisition. Instead, we hope that our contributions will inspire further research in the field of AP acquisition in individuals past their critical period.

Conclusion

This project gave us great insight into current literature into critical periods and neural networks. Especially it highlighted the fact that much about critical periods is still unknown and that there is still much more research needed to fill the knowledge gap. We learned that there is also great difficulty in studying brain activity due to the limitations of current imaging techniques and the lack of knowledge about the workings of the brain. Furthermore, the project demonstrated the in-depth understanding needed to propose a well formed research idea. Most importantly, the project illustrated how a research idea can change drastically in its formulation into a proper research project due to changing literature environment, especially in study fields that change as rapidly as AI.

References

Baharloo, S., Johnston, P. A., Service, S. K., Gitschier, J., & Freimer, N. B. (1998). Absolute pitch: an approach for identification of genetic and nongenetic components. The American Journal of Human Genetics, 62(2), 224-231.

Beck, D. L. (2008, November 17). Absolute Pitch, Perfect Pitch, Relative Pitch, and other Musical Notes [Editorial]. American Academy of Audiology.

Bellone, C., & Lüscher, C. (2012). Drug-evoked plasticity: do addictive drugs reopen a critical period of postnatal synaptic development? Frontiers in Molecular Neuroscience, 5, 1-7.

Bermudez, P. (2008). The Neural Correlates of Absolute Pitch (Order No. NR53300). Dissertation. McGill University (Canada).

Brown, Sachs, Cammuso & Folstein. Music Perception: An Interdisciplinary Journal, 21(1), 105-110.

Deutsch, D., Dooley, K., Henthorn, T., & Head, B. (2009). Absolute pitch among students in an American music conservatory: Association with tone language fluency. The Journal of the Acoustical Society of America, 125(4), 2398-2403.

Deutsch, D., Henthorn, T., Marvin, E., & Xu, H. (2006). Absolute pitch among American and Chinese conservatory students: Prevalence differences, and evidence for a speech-related critical period. The Journal of the Acoustical Society of America, 119(2), 719-722.

Drugs.com. (2017, November 1). Valproate Injection: Indications, Side Effects, Warnings. Retrieved November 27, 2017, from https://www.drugs.com/cdi/valproate-injection.html

Elite Music. (2013, November 21). Contemporary vs Classical Piano Lessons: Which is Better? Retrieved November 28, 2017, from https://elitemusic.ca/contemporary-vs-classical-piano-lessons-which-is-better/

Gervain, J., Vines, B. W., Chen, L. M., Seo, R. J., Hensch, T. K., Werker, J. F., & Young, A. H. (2013). Valproate reopens critical-period learning of absolute pitch. Frontiers in Systems Neuroscience, 7, 1-11.

Hensch, T. K. (2004). Critical period regulation. Annual Review of Neuroscience, 27(1), 549-579. Hensch, T. K., Bilimoria P. (2012). Re-opening Windows: Manipulating Critical Periods for Brain Development. Cerebrum 2012 (11), 1-18.

Hensch, T. K., Fagiolini, M., Mataga, N., Stryker, M. P., Baekkeskov, S., & Kash, S. F. (1998). Local GABA circuit control of experience-dependent plasticity in developing visual cortex.Science, 282(5393), 1504-1508. Hirose H., Kubota M., Kimura I., Ohsawa M., Yumoto M., Sakakihara Y. (2002). People with absolute pitch process tones with producing P300. Neuroscience Letters, 330, 247–250.

Hou, J., Chen, A. C., Song, B., Sun, C., & Beauchaine, T. P. (2016). Neural correlates of absolute pitch: A review. Musicae Scientiae, 21(3), 287-302.

Hübener, M., & Bonhoeffer, T. (2014). Neuronal plasticity: Beyond the critical period. Cell, 159(4), 727-737.

Ihalainen, T., Kuusela, L., Turunen, S., Heikkinen, S., Savolainen, S., & Sipilä, O. (2014). Data quality in fMRI and simultaneous EEG–fMRI. Magnetic Resonance Materials in Physics, Biology and Medicine, 28(1), 23-31.

Levitin, D. J., & Rogers, S. E. (2005). Absolute pitch: perception, coding, and controversies. Trends in cognitive sciences, 9(1), 26-33.

Levitin, D. J., & Zatorre, R. J. (2003). On the nature of early music training and absolute pitch: A reply to brown, sachs, cammuso, and folstein. Music Perception: An Interdisciplinary Journal, 21(1), 105-110.

Lewis, C. M., Bosman, C. A., & Fries, P. (2015). Recording of brain activity across spatial scales. Current Opinion in Neurobiology, 32, 68-77.

Lillard, A. S., & Erisir, A. (2011). Old dogs learning new tricks: Neuroplasticity beyond the juvenile period. Developmental Review, 31(4), 207.

Lockhead, G. R., & Byrd, R. (1981). Practically perfect pitch [Abstract]. The Journal of the Acoustical Society of America, 70(2), 387-389.

Loui, P., Li, H. C., Hohmann, A., & Schlaug, G. (2011). Enhanced cortical connectivity in absolute pitch musicians: a model for local hyperconnectivity. Journal of cognitive neuroscience, 23(4), 1015-1026.

Notter, M. (2016). Introduction to Neuroimaging. Retrieved November 28, 2017, from http://miykael.github.io/nipype-beginner-s-guide/neuroimaging.html

Oldfield, R. C. (1971). The Assessment and Analysis of Handedness: The Edinburgh Inventory. Neuropsychologia, 9, 97-113.

Panzeri, S., Brunel, N., Logothetis, N. K., & Kayser, C. (2010). Sensory neural codes using multiplexed temporal scales. Trends in Neurosciences, 33(3), 111-120.

Petrov, A. Y., Herbst, M., & Stenger, V. A. (2017). Improving temporal resolution in fMRI using a 3D spiral acquisition and low rank plus sparse (L+S) reconstruction. Neuroimage, 157, 660.

Preston, D. C. (2006, November 30). MRI Basics. Retrieved November 28, 2017, from http://casemed.case.edu/clerkships/neurology/Web%20Neurorad/MRI%20Basics.htm

Profita, J., Bidder, T. G., Optiz, J. M., & Reynolds, J. F. (1988). Perfect pitch. American Journal of Medical Genetics Part A, 29(4), 763-771.

Pylkkanen, Liina. (2016). Neural Bases of Language & Neurolinguistics Lecture [PowerPoint slides]. Retrieved from http://www.psych.nyu.edu/pylkkanen/NeurolinguisticsFall2016/Lectures/04.pdf

Saffran, Jenny R. (2003) Absolute pitch in infancy and adulthood: the role of tonal structure. Developmental Science, 6(1), 35–43.

Schlaug, G., Jäncke, L., Huang, Y., Staiger, J. F., & Steinmetz, H. (1995). Increased corpus callosum size in musicians. Neuropsychologia, 33(8), 1047-1055.

Schulze, K., Gaab, N., & Schlaug, G. (2009). Perceiving pitch absolutely: Comparing absolute and relative pitch possessors in a pitch memory task. BMC Neuroscience, 10(1), 106-118.

Schwenzer, M., & Mathiak, K. (2011). Numeric aspects in pitch identification: an fMRI study. BMC Neuroscience, 12(1), 26-35.

Takeuchi, A. H., & Hulse, S. H. (1993). Absolute pitch. Psychological bulletin, 113(2), 345.

Werker, Janet F., Hensch, Takao K. (2015). Critical Periods in Speech Perception: New Directions. Annual Review of Psychology 66, 173-196.

Zatorre, R. J. (2003). Absolute pitch: a model for understanding the influence of genes and development on neural and cognitive function. Nature Neuroscience, 6(7), 692-695.

Snoek, J., Larochelle, H., & Adams, R. P. (2012). Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems (pp. 2951-2959).

Sabour, S., Frosst, N., & Hinton, G. E. (2017). Dynamic Routing Between Capsules. In Advances in Neural Information Processing Systems (pp. 3857-3867).

Arxiv, E. (2017). The Revolutionary Technique That Quietly Changed Machine Vision Forever. MIT Technology Review. Retrieved 28 November 2017, from https://www.technologyreview.com/s/530561/the-revolutionary-technique-that-quietly-changed-machine-vision-forever/ Historic Achievement: Microsoft researchers reach human parity in conversational speech recognition - The AI Blog. (2017). The AI Blog. Retrieved 28 November 2017, from https://blogs.microsoft.com/ai/2016/10/18/historic-achievement-microsoft-researchers-reach-human-parity-conversational-speech-recognition/ MNIST handwritten digit database, Yann LeCun, Corinna Cortes and Chris Burges. (2017). Yann.lecun.com. Retrieved 29 November 2017, from http://yann.lecun.com/exdb/mnist/ This AI Technique Was Kept Quiet so Spammers Wouldn’t Misuse It. (2017). MIT Technology Review. Retrieved 28 November 2017, from https://www.technologyreview.com/the-download/609228/this-ai-technique-was-kept-quiet-so-spammers-wouldnt-misuse-it/