Experiments 😎
I believe in the knack of going into the depths. To understand. To dissect things apart until you can confidently say, "I know how it works".
​
In my free time, I experiment. This is separate from the projects that I have professionally undertaken. This is something I do for fun. For pure fun.
Sentence Transformers
When I read through the paper of Sentence Transformers, What striked my mind was, How did cross entropy loss work in bringing the embeddings closer in the vector space!
​
I recreated the research paper. I basically took a non-fine tuned transformers encoder. Fine tuned it with 2 different kinds of loss functions.
Contrastive loss
Cross entropy loss
​
This helped me validate the results that Cross Entropy despite being very unintuitive for this particular problem statement, surprisingly worked well!
General Purpose T5 Model
Further fueled by my interests in understanding how an encoder-decoder model can be adapted for multi-purpose tasks, I trained a base T5 model on multiple different tasks which includes but not limited to Abstractive Summarization, Language Translation, Grammar Correction. I prepended the encoder part of the model with a special unique token for each task.
​
I found it to work surprisingly well and with decent accuracy considering the limited dataset and the time I trained it for.
Understanding Emotions purely from Audio waves
Understanding emotions from text is easy. It's well explored and well tested. How about exploring what the person is feeling purely from the audio waves ? I tried this by extracting the Mel-Spectrogram (which is basically the audio frequency). I used basic CNN models to identify the prospect of rightly classifying someones emotions from their speech!
​
Did it work ?
For high octane emotions, IT DID! It was able to pick out speech tones of anger, extreme happiness, or tones of visible sadness. The model suffered when the emotions were more nuanced like calmness, happiness and so on. Nevertheless, It was an extremely interesting and fun one I did!
