Below is a reminder of our two previously proposed solutions:
- Music class with NIRS - participation & listening based (singing & instruments)
- It is financially-wise, considering that our budget is limited and we want to test on large enough samples to get enough data for our computational model
- It doesn’t make any noise, so we make sure that young individuals involved in our experiment are actually responding to “music” instead of random noise.
- Subjects can move around, which is great since our test subjects are toddlers
- It can be used with babies
- A setting of music class makes it easier for young individuals to relax and provide brain data similar to real-life situations
- It enables us to see how they respond behaviorally and neuro-biologocially to different musical elements
- It makes it easier for us to build connections between music pieces and different music elements involved in each music piece
2. Computational Machine Learning Model:
How it works:
a) Input :
- brain data for each individual (specific regions with impairment)
- corresponding “music” data for each individual (for each musical element, whether or not this individual can respond to)
- To find out the specific correspondence between brain activities and musical elements/types.
- To find out what musical elements each music piece used in our experiment is composed of.
- To find out a specific therapy for each individual with SLI by deriving a best combination of those musical elements that they should be exposed to according to our model.
- For any arbitrary but particular individual with SLI, we provide a musical therapy that is suggested by our model.
Why the computational machine learning model:
- It utilizes all the data we get in an optimal way.
- It helps to show the connection of data that is otherwise very hard to figure out
- It gives us the currently best music therapy that we can provide to any young individual involved in our experiment. Also, once we start applying this musical therapy with the individual, and we get more data, we can then feed new data into our model and obtain more accurately predicted solutions along the way.
3 “yes, ands…” for solution 1(music class with NIRS):
- Implementing music in real life situations instead of just in a music class environment
Response: Instead of just having babies sit in a room and listen to music together, we are going to have the young children do language tasks both silently and while listening to different types of music. We think that this will help narrow down what music affects which part of the brain, and help us figure out how to help each child individually.
2. Comparing the effects of melody and lyrics
Response: Yes, it is important that we compare different types of music to try to piece together which musical elements have certain effects. We will compare both vocal and instrumental music, in the child’s native language and a language they do not know, and music that is more percussive than instrumental. We hope to test which of these musical factors helps the child improve.
3. Singing may be more effective than learning instruments
Response: Agreed. Adding an element of learning to play instruments may overcomplicate our solution. We plan to do more listening, along with language tasks.
Summary: Yes, music class is a good idea - it would be good to focus on singing and implement the collected data into a machine learning model. There were concerns with certain factors of the music class idea, such as having kids play instruments and how music class would translate to helping kids in real life situations. We have focused our idea to include the best parts of both solutions, and have cut out some of the variations, like the instruments. Hopefully, this will lead to more straightforward results.
3 “yes, buts…” for solution 1(music class with NIRS):
- Why NIRS? Seems broad
Response: NIRS is the best option for us because we do not want to limit how much the children can move and don’t need the specificity of an MRI or something similar. NIRS is also relatively inexpensive.
2. More detection than a cure
Response: Without a next step, this is true. So we have decided to use the data gathered from the NIRS in our music class in a machine learning model. We hope that this combined solution will give another option to kids and families struggling with language learning disabilities.
3. Find a way to utilize the data from NIRS
Response: Good idea! As stated above, this was missing in our first draft. We have decided to use the NIRS data in our machine learning model.
Summary: Generally, people think that music class is a good idea, but it may be too broad and seems incomplete without a way to utilize the data from NIRS. To make our solution better, we have found a way to utilize the NIRS data. This will help us track the relationship between certain musical elements and certain language deficits. Hopefully with this new solution, we can help lots of kids with SLI.
3 “yes, ands…” for solution 2(computational machine learning model):
- “Exposing patients to different music types would help even more and seems like more of a real solution”
Response: Yeah, this is exactly what we are thinking about. So we do plan to expose patients to a variety of music types, and this is not only for the purpose of a potentially working musical therapy but also and more for the purpose of understanding how those patients respond to different music types. A lot of relevant research in the field has suggested that there exists an exact correspondence between missing musical behaviors and missing linguistics behaviors; however, it is yet to be discovered which musical behaviors are corresponding to which linguistics behaviors. Thus, we hope to get closer to the answer through our research.
2. “Research potential musical combinations that are best for therapy, and check if there has already been existing data.”
Response: This is a very good suggestion! We definitely need to look into what are potentially good music for therapy so that we can use them when we do our second test - exposing SLI patients and a control group to different music types and see which type of music activates which part of the brain, since we never want to use some music that is not likely to activate any essential brain part.
3. (A lot of people) like this solution better than our first solution.
Response: We are really happy that people like our idea of applying machine learning model. In fact, we are planning to incorporate our first solution as part of the way to gather data for this machine learning model. Please refer to our revised solution for more details.
Summary: We have got many suggestions on how we should design our experiment for testing the correspondence between different types of music and different parts of the brain. Based on the suggestion, we plan to look for those types of music that are potentially good musical therapy in advance, but also make sure that we do provide a wide variety of music for the experiment. In this way, we try to cover as thorough data as possible but also make sure we mainly focus on the more relevant music without wasting time on the absolutely non-relevant ones.
3 “yes, buts…” for solution 2(computational machine learning model):
- “It will be really difficult to make specified models for each individual. That will take so much time, money and research.”
Response: It seems that there’s some confusion in what machine learning is and how a machine learning model works. So let us try to explain it better. Say after the experiment, we have got some brain data and we have some other music data, and there is some correlation but we need this model to figure this correlation out. Therefore, we plug in part of the data into the model and after it runs, it will tell us what it thinks is the correlation. Therefore, next time, when we have a new patient and their brain data, we can plug this brain data in and the model will automatically output a corresponding music data (which is in this case a music therapy). So the idea of machine learning is that we develop a general model, but we plug some inputs we get an output that is specific to that input. And in fact, the part of machine learning does not cost much, it’s basically using existing data modes and coding. The costly and time-consuming part would probably be how we get the data, which is the experiment part.
2. “What is the combinational solution that the machine learning will bring about? I feel like anyone can say that we want to customize solution by utilizing a program. It would be better if there’s any relevant research on the specific steps that the program will employ.”
Response: In fact, machine learning is indeed this easy. Anyone can apply machine learning on any data and predict any relevant feature to their data. It is just a general technique for data analysis and prediction. So you can think of the way it provides specific result is like developing a mapping, given an input, the model will output the corresponding output that is specific to that input.
3. How will this impact kids (SRI patients) who are not musically inclined?
Response: Thanks for bringing this up. In fact, the reason we want to focus on musical approach to tackle SRI is that a lot of relevant research in the field has suggested that there exists an exact correspondence between missing musical behaviors and missing linguistics behaviors, particularly SRI. The only thing is that it is yet to be discovered which musical behaviors are corresponding to which linguistics behaviors, but hopefully we can manage to get closer to the answer through our research.
Summary: It seems that people are still confused about machine learning in general, and the most common concern is how machine learning model provides specific/customized solution. So a very high-level understanding of machine learning is like: we have a bunch of data for X1, X2, X3,...,XN, and a bunch of data for Y1,Y2,Y3,...,YN, and now we want to see how they are related. We basically just throw all these data into the model, and it will generate a sort-of best fitting relationship between all those features. So how the exact relationship looks like might be complicated, but for now don’t worry about it. Just think of it as a function that maps X to Y, so next time, when we get a new X, and we want to know what might be a possible value for Y. We just plug in X to the model and it will tell us what it thinks Y might be.
Similarly here, after conducting the experiment, we will have brain data of a lot of individuals and we have data how different types music are activating the brain for all those individuals as well. The machine learning model can tell us what it think the relationship between these two groups of data. Now, with this model, whenever a new patient comes, we can just plug the brain data of that new patient into the model and it will output the corresponding combination of music as the therapy.
Our revised solution:
Our revised new solution is a combination of our two previously proposed solutions. This will allow us to have both a method for detection of the dysfunctional brain area as well as a way to interpret and process our feedback. We changed our solutions based on the feedback we received to create more specific and clear methods for our work. Rather than gathering the data using NIRs for our initial detection during a music class we think that having the subjects perform language tasks with NIRs is much more translational and relevant to the SLI disorder. This will provide feedback about what specific areas of their brains are not functioning properly, whether that be no activation at all or suboptimal activation. Then, we will utilize different types of music to see how the different musical elements activate the brain and what can activate the dysfunctional brain areas. The machine learning model will be important and utilized two separate times for our solution. First, to build a connection between linguistics and the brain behavior from the language tasks. Then, again to after exposure to music to draw connections between the music types and brain behavior. Overall, we will be able to generate a more well-rounded solution to the problem as we will be able to both detect the problem and use the computational model to solve the detected problem.