BEGIN:VCALENDAR VERSION:2.0 PRODID:IEEE vTools.Events//EN CALSCALE:GREGORIAN BEGIN:VTIMEZONE TZID:America/New_York BEGIN:DAYLIGHT DTSTART:20260308T030000 TZOFFSETFROM:-0500 TZOFFSETTO:-0400 RRULE:FREQ=YEARLY;BYDAY=2SU;BYMONTH=3 TZNAME:EDT END:DAYLIGHT BEGIN:STANDARD DTSTART:20261101T010000 TZOFFSETFROM:-0400 TZOFFSETTO:-0500 RRULE:FREQ=YEARLY;BYDAY=1SU;BYMONTH=11 TZNAME:EST END:STANDARD END:VTIMEZONE BEGIN:VEVENT DTSTAMP:20260929T183433Z UID:FF5DE119-CAD8-4FD5-A573-E71936830E31 DTSTART;TZID=America/New_York:20261024T100000 DTEND;TZID=America/New_York:20261024T110000 DESCRIPTION:Abstract:\n\nAs deep learning models scale into multi-billion p arameter spaces\, modern enterprise architectures face an immediate barrie r: the memory bandwidth wall. Standard operator execution patterns bot- tl eneck high-performance systems by forcing intermediate tensor states to re peatedly cycle through high-bandwidth memory (HBM). This lecture addresses the mathematical and hardware codesign patterns required to achieve subli near search mandates and drive maximum runtime execution throughput across real-time\, multi-modal environments.\n\nThrough a deep dive into PyTorch implementation mechanics\, we examine the dual paradigms of data vs. comp ute vectorization. We analyze how transforming unstructured real-world ent ities into continuous\, high-dimensional dense numeric representations pro vides critical topological con- text for downstream similarity checks. Mov ing past standard eager-mode execution patterns\, we demonstrate how custo m OpenAI Triton and CUDA fused memory kernels mathematically com- bine sca le\, mask\, and softmax routines into localized execution blocks inside fa st SRAM registers\, bypassing HBM (High Bandwidth Memory) latency loops co mpletely.\n\nFurthermore\, this session explores advanced index topologies |speci cally Hierarchical Navi- gable Small Worlds (HNSW) coupled with V ector Quantization (VQ) boundaries. By applying Vector Quantization to com press vast oating-point vector spaces into low-bit indices\, we illustrate how to bridge the gap between continuous neural state spaces and discrete tool selection loops. Finally\, we synthesize these optimization mechanic s within the framework of Agentic AI control loops (ReAct- Reasoning and A cting)\, providing a uni ed approach to cognitive scaling\, multi-agent synchronization\, and real-world hallucination mitigation\n\nhttps://event s.vtools.ieee.org/m/575616\n\nSpeaker(s): Dr. Baris Kazar\, \n\nRochester\ , New York\, United States\, Virtual: https://events.vtools.ieee.org/m/580 546 LOCATION:Rochester\, New York\, United States\, Virtual: https://events.vto ols.ieee.org/m/580546 ORGANIZER:mbkazar@gmail.com SEQUENCE:18 SUMMARY:{AI Talks with Tea&Coffee #53 @ RIT} : Advanced Latent Representati ons and Computational Optimization: A Summary of PyTorch Experiences in De ep Architectures\, RAG\, and Agentic AI: Vector Embeddings\, Vectorization and Vector Quantization URL;VALUE=URI:https://events.vtools.ieee.org/m/580546 X-ALT-DESC:Description: <br /><p data-removefontsize="true" data-originalco mputedfontsize="16">Abstract:</p>\n<p data-removefontsize="true" data-orig inalcomputedfontsize="16">As deep learning models scale into multi-billion parameter spaces\, modern enterprise architectures face an immediate barr ier: the memory bandwidth wall. Standard operator execution patterns bot- tleneck high-performance systems by forcing intermediate tensor states to repeatedly cycle through high-bandwidth memory (HBM). This lecture address es the mathematical and hardware codesign patterns required to achieve sub linear search mandates and drive maximum runtime execution throughput acro ss real-time\, multi-modal environments.</p>\n<p data-removefontsize="true " data-originalcomputedfontsize="16">&nbsp\;</p>\n<p data-removefontsize=" true" data-originalcomputedfontsize="16">Through a deep dive into PyTorch implementation mechanics\, we examine the dual paradigms of data vs. compu te vectorization. We analyze how transforming unstructured real-world enti ties into continuous\, high-dimensional dense numeric representations prov ides critical topological con- text for downstream similarity checks. Movi ng past standard eager-mode execution patterns\, we demonstrate how custom OpenAI Triton and CUDA fused memory kernels mathematically com- bine scal e\, mask\, and softmax routines into localized execution blocks inside fas t SRAM registers\, bypassing HBM (High Bandwidth Memory) latency loops com pletely.</p>\n<p data-removefontsize="true" data-originalcomputedfontsize= "16">&nbsp\;</p>\n<p data-removefontsize="true" data-originalcomputedfonts ize="16">Furthermore\, this session explores advanced index topologies|spe ci cally Hierarchical Navi- gable Small Worlds (HNSW) coupled with Vecto r Quantization (VQ) boundaries. By applying Vector Quantization to compres s vast&nbsp\; oating-point vector spaces into low-bit indices\, we illustr ate how to bridge the gap between continuous neural state spaces and discr ete tool selection loops. Finally\, we synthesize these optimization mecha nics within the framework of Agentic AI control loops (ReAct- Reasoning an d Acting)\, providing a uni ed approach to cognitive scaling\, multi-age nt synchronization\, and real-world hallucination mitigation</p>\n<p data- removefontsize="true" data-originalcomputedfontsize="16">&nbsp\;</p>\n<p d ata-removefontsize="true" data-originalcomputedfontsize="16">https://event s.vtools.ieee.org/m/575616</p> END:VEVENT END:VCALENDAR