Ignacio
Fresneda Gallardo

Applying PCMCI+ to discover causal links between five environmental variables (PM2.5, O3, temperature, humidity, physical activity) and respiratory rate in paediatric asthma patients in Delhi. Extends the merged DAPHNE 1.0+1.5 cohort, adds pollution source classification and validates the pipeline on an adult London cohort (INHALE).

Speaker-level dysphonia detection using HuBERT embeddings. Comparison between hierarchical aggregation and simple pooling strategies over 1,123 speakers from the Saarbrücken Voice Database.

Custom Triton/CUDA kernels for RMSNorm, LayerNorm, GELU, SiLU, attention, RoPE and FlashAttention on the GLM-ASR speech-to-text model. Zero quality degradation.

Daily song ranking prediction over 467K entries. LSTM with a 90-day window outperformed XGBoost, MLP and linear regression. K-means clustering revealed seasonal patterns in global music tastes.

Recurrent network (firing rate → Leaky Integrate-and-Fire) trained with reinforcement learning on a binary-choice task inspired by primate experiments. Transfer learning with biologically realistic behaviour.

Analysis of 11,113 brain activity trials in primates (Romo et al., 1997). Identification of 79 significantly tuned neurons out of 560 and quantification of contraction bias in short-term memory.