Atharv Sharma atharvgaur1845

_{FRONT MATTER}

Second-year CS undergraduate at BITS Pilani, working toward frontier AI research in representation learning and world models.

The method is fixed even when the subject isn't: read the paper, rebuild it from scratch, reproduce the result, find the hidden assumptions — then look for where they break. Everything in this repository exists to run that loop faster.

_{mathematics → theory → paper → implementation → reproduction → experiment → failure analysis → original ideas}

_COORDINATES

field   representation learning,         world models lab     repositories site    atharvgaur1845.github.io in      atharv-gaur x       @tarkrishi

_{values  evidence over hype ·
        depth over breadth ·
        originality over replication}

`01 · current investigation`

_FOCUS	temporal JEPA for satellite time-series
_STATUS	building — training scaffold up, instrumenting rollouts
_SINCE	2026-06-30
_REPO	multi_temporal_jepa
_HYPOTHESIS	Latent prediction without reconstruction keeps the information that matters for downstream temporal reasoning — and its failure mode (collapse or divergence) is measurable long before accuracy metrics notice.
_LATEST	Multi-temporal JEPA scaffold trains end-to-end on satellite sequences. Next measurement: how prediction error in latent space grows with rollout horizon, per spectral band.

_{auto · research/current.md · rebuilt on push + daily}

`02 · lab notebook`

`2026-07-21`	rollout instrumentation for multi-temporal JEPA _#experiment Wired divergence logging into the training loop so latent prediction error is tracked per horizon step, per band, at every checkpoint — not computed after the fact. If the hypothesis is that collapse announces itself early, the instrument has to be running from step zero.
`2026-07-11`	predictor capacity is not free _{#experiment #failure} Doubling predictor depth looked like a clean win on one-step prediction and quietly wrecked long-horizon rollouts — the predictor started memorizing trajectory quirks instead of dynamics. Lesson recorded: evaluate at the horizon you care about, never at the horizon that is cheap.
`2026-07-02`	why JEPA over masked autoencoding, in one paragraph _#reading Reconstruction objectives spend capacity on exactly the pixels I don't care about (clouds, sensor noise); predicting in latent space lets the encoder discard them. The open question I keep circling: without reconstruction, what stops the latent from discarding too much? That question is the project.

_{auto · last 3 of research/notebook.md · full notebook →}

`03 · activity trace`

`2026-10-12`	Boundary_book	_{feed: refresh 2026-10-12T10:26Z [skip ci]}
`2026-09-10`	CUHK-X-CompetitionSmallModelTrack	_{EXP-148: seed soup is decisively NEGATIVE (-2.39, 4/4 folds) and underc…}
`2026-08-26`	multi_temporal_jepa	_{Recorded in the runbook, and 01_stage_pastis.sh now exports the overrid…}
`2026-08-17`	flow_game	_{Replace room-timer death with overtime across all three builds}
`2026-08-05`	NeuralDebrisRemovalinStreakDetectionModels	_{Update README}

_{auto · github api · newest repository: ChronoSched (created 2026-08-01)}

`04 · research program`

_{OPEN QUESTIONS — the standing ones, revised as evidence comes in}

why do learned representations fail — and can the failure be seen coming?
how stable are latent spaces over long prediction horizons?
which assumptions in modern self-supervised learning break first?
how far can predictive learning go without reconstruction?
can world models generalize beyond observed trajectories?

`05 · experiments`

`multi_temporal_jepa · flagship · active` _{repository →}

_WHY	Test whether latent prediction — no reconstruction — retains what matters in satellite time-series, and where it collapses over long horizons.
_RESULT	Working training scaffold for multi-temporal JEPA on satellite sequences; currently instrumenting rollout-divergence measurements.
_{HARD PARTS}	Preventing representation collapse without contrastive negatives, and deciding what "good" latents even means when there are no labels to check.
_NEXT	Ablate predictor depth vs. horizon length; log divergence curves per band.

`industrial time-series representation learning · internship · paused` _{repository →}

_WHY	Can self-supervised encoders separate machine fault modes from raw vibration signals, without a single label?
_RESULT	VAE-GAN / VQ-GAN encoders on the CWRU bearing dataset; spectral preprocessing pipeline in a companion repo (timeseries-fft-matching).
_{HARD PARTS}	Vibration data punishes sloppy windowing — spectral leakage quietly destroys everything downstream.
_NEXT	Revisit with a JEPA-style objective instead of generative reconstruction.

`foundations, rebuilt by hand · practice · archived` _{repository →}
_WHY	You don't own a method until you've implemented it from nothing.
_RESULT	Linear/logistic regression, k-means compression, CNNs, U-Net segmentation and anomaly detection — each written from first principles before touching a framework.
_{HARD PARTS}	Resisting the urge to import the answer.
_NEXT	The series continues with attention and diffusion, one primitive at a time.

_{auto · projects.yml · plates drawn by hand in assets/lab/}

`06 · reading pipeline`

`◆ implementing`	I-JEPA: Self-Supervised Learning from Images with a Joint-Embedding Predictive… _{Assran et al. · CVPR 2023}	_{the blueprint for multi_temporal_jepa; predictor design carries over, temporal axis does not}
`● annotated`	V-JEPA: Video Joint-Embedding Predictive Architecture _{Bardes et al. · 2024}	_{masking strategy over time is the interesting part; satellite revisit gaps break their assumptions}
`● annotated`	A Path Towards Autonomous Machine Intelligence _{LeCun · 2022}	_{the position paper behind the program; hierarchical world models section re-read quarterly}
`▸ reading`	On the Duality Between Contrastive and Non-Contrastive Self-Supervised Learning _{Garrido et al. · ICLR 2023}	_{the collapse question, made precise — directly relevant to training without negatives}
`● annotated`	VICReg: Variance-Invariance-Covariance Regularization _{Bardes et al. · ICLR 2022}	_{variance term as anti-collapse mechanism; candidate diagnostic for the seismograph idea}
`○ queued`	DreamerV3: Mastering Diverse Domains through World Models _{Hafner et al. · 2023}	_{world models with actions — where the program goes after passive prediction}

_{auto · papers/reading.md · ◆ implementing · ● annotated · ▸ reading · ○ queued}

`07 · failure log`

_{negative results are results. entries tagged #failure in the notebook land here on their own.}

2026-07-11

predictor capacity is not free
Doubling predictor depth looked like a clean win on one-step prediction and quietly wrecked long-horizon rollouts — the predictor started memorizing trajectory quirks instead of dynamics. Lesson recorded: evaluate at the horizon you care about, never at the horizon that is cheap.

_{auto · entries tagged #failure in research/notebook.md}

`08 · idea incubator`

_[growing]	measure latent-space stability as an early-warning signal for representation collapse — a "seismograph" for training runs
_[growing]	JEPA-style objective for industrial vibration data, replacing the generative CWRU pipeline
_[seed]	does temporal irregularity (uneven satellite revisit times) demand time-aware positional structure in the predictor?
_[seed]	spectral methods as a bridge between graph structure and temporal structure — same eigenproblem, two fields
_[seed]	a benchmark for "assumption breakage": perturb the assumption a method relies on, measure graceful vs. catastrophic degradation

_{auto · research/ideas.md · seed → growing → ready; ready ideas graduate into experiments}

`09 · instruments`

_LANGUAGES	`python · c++ · java · c`
_LEARNING	`pytorch · tensorflow · numpy · pandas · scikit-learn`
_VISION	`transformers · cnns · gans · gnns · opencv`
_INFRA	`linux · git · docker · cuda · jupyter`
_SIMULATION	`ros2 · gazebo`

_{this profile is a generated artifact — the readme is never edited by hand.}
_{layout templates/lab.md · data research/ papers/ projects.yml + github api · builder build_readme.py via build.yml (on push + daily) · last build 2026-10-12 13:09 utc}
_{how it works: architecture · design system · the three concepts}

Navigation Menu

Atharv Sharma atharvgaur1845

Achievements