The AI @ ADI Reading Digest: Recommended papers and research about AI at the edge.

📚 The AI @ ADI Reading Digest: Recommended paper and research about AI at the edge.

At ADI, AI means developing world-class models that overcome latency, power, and cost constraints to deliver real-world insights. June brings its set of new ideas and practical breakthroughs, as the field continues to push toward more capable, efficient, and deployable systems. This month’s selections reflect that momentum, with methods that stretch model capability, improve efficiency, and move AI closer to real-world impact at the edge. Whether you’re squeezing in a quick read or diving deeper, here’s our list of recommended papers to explore.

📄 Does YOLO Really Need to See Every Training Image in Every Epoch?
👍🏼 Recommended by: Debanjan Ghosh
💡 Why it’s worth reading: This paper reframes a hidden inefficiency in YOLO training—blindly iterating over all data each epoch—and replaces it with a simple but powerful curriculum-style sampler (AFSS) that prioritizes hard examples while sparsely revisiting “learned” ones. As highlighted in the presentation, the elegance lies in its practicality: a periodic precision/recall-based difficulty score drives 2%/40%/100% sampling without architectural changes, yielding ~1.4–1.6× speedups and better accuracy. It connects cleanly to curriculum learning, replay, and importance sampling—suggesting broader gains for data-efficient training across edge and robotics workloads.

📄 Universal Redundancies in Time Series Foundation Models
👍🏼 Sambarta Dasgupta 
💡 Why it’s worth reading: This paper brings mechanistic interpretability into the largely unexplored space of time-series foundation models (TSFMs), revealing a surprisingly strong result: large portions of these transformer stacks are structurally redundant. Across models and datasets, entire layers and significant fractions of attention heads can be ablated with minimal performance loss, implying TSFMs behave more like overparameterized kernel regressors than tightly coupled deep models. The accompanying presentation emphasizes how these redundancies are not just inefficiencies but explanatory—pinpointing heads that drive pathologies like motif copying and seasonality bias—suggesting a clear path toward principled pruning, interpretability, and edge-efficient forecasting systems.

📄 BYOL for Audio: Self-Supervised Learning for General-Purpose Audio Representation
👍🏼 Recommended by: 💡 This 2021 paper from researchers at NTT Corporation in Japan applies the bootstrap-your-own-latent (BYOL) approach to train an audio representation model. In BYOL, the student and teacher get different augmentations (views) of the input, and the student learns to match the teacher, while the teacher's weights are EMA of the student's weights. The innovation here is the data augmentation scheme to apply BYOL to audio domain.