Research
I am interested in building an intelligent system that is evolving in nature and which reuses its current knowledge to learn new tasks. My current research revolves around building neural training techniques that work better than Backpropagation across domains.
|
|
Neuro-mimetic Task-free Unsupervised Online Learning with Continual Self-Organizing Maps
Hitesh Vaidya,
Travis Desell,
Ankur Mali,
Alexander Ororbia
Preprint
pdf  / 
arxiv  / 
bibtex
CSOM, a competitive learning based neural model that reduces catastrophic forgetting in a truly unsupervised online class incremental learning setting. We show with a theoritical proof that CSOM converges steadily to a stable model that stores linearly separable clusters of input samples without needing any class labels or task boundaries.
|
|
Reducing Catastrophic Forgetting in Self Organizing Maps with Internally-Induced Generative Replay (Student Abstract)
Hitesh Vaidya,
Travis Desell,
Alexander Ororbia
AAAI-22 Student Abstract and Poster Program
pdf  / 
arxiv  / 
bibtex
In this work, we propose the c-SOM (continual SOM), an SOM that actively reduces the amount of forgetting it experiences. Specifically, we modify the SOM’s decay to be task-dependent and extend its units to self-induce a form of generative rehearsal to improve memory retention.
|
|
Stream-51: Streaming Classification and Novelty Detection From Videos
Ryne Roady,
Tyler Hayes,
Hitesh Vaidya,
Christopher Kanan
CVPR Workshop on Continual Learning, 2020
Video  / 
code  / 
bibtex
In this work, we introduce Stream-51, a new dataset for streaming classification consisting of temporally correlated images from 51 distinct object categories and additional evaluation classes outside of the training distribution to test novelty recognition. We establish unique evaluation protocols, experimental metrics, and baselines for our dataset in the streaming paradigm.
|
|
Reducing Catastrophic Forgetting in Self-Organizing Maps
Hitesh Vaidya
pdf  / 
bibtex
This is my MS Thesis in which I provide a survey about lifelong/continual learning. I mention about current methods, applications and limitations in lifelong learning. Further, I show that unsupervised learning approaches like self-organizing maps (SOMs) can be used as oracle to perform replay in lifelong learning neural networks. In the process, I display several novel methods to reduce forgetting in class incrementally trained SOMs and provide relevant baselines.
|
About me
My life in Rochester, NY was really memorable because my friends.
I like exploring new things in life and road trips are one of the best ways to do that in my opinion. Badminton and running are my all time favorites besides swimming and weight training.
When indoors, I like doing Yoga, reading and listening to good music (I have a wide range of genres from Indian classical to Pop, R&B). I can play a bit of Tabla and flute too.
|
|