Principal Scientist - Applied Research, ASML (Multimodal Foundation Models)

About Us: Adobe Firefly is the industry's premier generative AI model family, built to be safe for commercial use and deeply embedded into the workflows of millions of creators. The Opportunity: We are looking for a Principal Scientist of Applied Research for our Applied Science and Machine Learning (ASML) in India. This organization is responsible for scaling, quality, and innovation in multimodal training data —spanning image, video, and audio —that powers Adobe Firefly’s foundational generative and editing models . The Multimodal Data Research organization sets the data foundation for Firefly’s multimodal intelligence, owning first‑party strategy and execution across ingestion, filtering, captioning, generation and editing‑centric datasets, data‑driven training experiments, and next‑generation data frameworks. By grounding model training in high‑quality, diverse, and workflow‑authentic data, the team directly builds Firefly’s creative capabilities. What You’ll Do: Multimodal Model Training: Guide the Applied Research team on architecture and training strategies for models that bridge text, pixels, temporal frames, and audio waveforms. Cross-Modal Data Ablations: Pioneer our approach to data ablations across diverse modalities. You will build rigorous experiments to understand how dataset composition, cross-modal alignment (e.g., text-to-video, image-to-audio), pruning, and blending impact final model capabilities and commercial viability. Research Leadership: Guide a world-class team of applied scientists and ML researchers, encouraging rapid experimentation, deep scientific rigor, and groundbreaking progress in multimodal representation learning. Scale & Infrastructure: Partner closely with Core Engineering and Infrastructure teams to optimize distributed training pipelines across massive GPU clusters, ensuring efficient resource utilization and training stability for complex multimodal workloads. Bridge Research to Product: Translate ground breaking AI research into production-ready foundation models that power Adobe's flagship tools (Premiere Pro, After Effects, Photoshop, GenStudio). Global Collaboration: Collaborate with global AI research leaders, data engineering teams, and legal/trust teams to ensure our models remain the industry standard for commercial safety and creative quality. What You Need to Succeed: Research & Engineering Leadership: 10 years of experience in Applied Machine Learning or AI Research, and at least 4 years mentoring and scaling high-performing research teams. Multimodal Foundation Model Expertise: Hands-on experience training models that process multiple modalities (e.g., Diffusion Transformers (DiT), Multimodal LLMs, joint embedding architectures like CLIP, or generative video/audio models) from scratch. Proficiency in Multimodal Data: Experience crafting and delivering data ablation studies. You understand how to balance diverse datasets, handle varying data ingestion rates, curate high-quality multimodal pairs, and manage multi-objective training. Distributed Training at Scale: Architectural understanding of training massive models using frameworks like PyTorch, Megatron-LM, DeepSpeed, or FSDP. You know how to troubleshoot loss spikes, gradients, and cluster bottlenecks specific to multimodal architectures. Scientific Rigor: Turning ambiguous research questions into structured experiments with measurable outcomes. Education: PhD or MS in Computer Science, Artificial Intelligence, Machine Learning, or a closely related highly quantitative field, or equivalent experience. Adobe is an equal opportunity employer. We hire dedicated individuals, regardless of gender, race, ethnicity, ancestry, age, disability, sexual orientation, gender identity or expression, veteran status, cultural background or religious beliefs. We know that when our employees feel appreciated and included, they can be more creative, innovative and successful. This is what it means to be Adobe For All. Lea ... (truncated, view full listing at source)