Explore

Slice the corpus on any axis.

Filter by firm, year, function, metric or slide category — then see the result as a narrative arc, a treemap of slide types, or a catalog of slides.

clear
51,416 slides across 1229 decks match.
Treemap Arc Catalog
Offence is the best defence
PwC · 2016 · key_messages
Offence is the best defence
78% — CEO support for FinTech integration
 4.2 Adopting a ‘mobile-first’ approach
PwC · 2016 · industry_trends
4.2 Adopting a ‘mobile-first’ approach
61% — Client mobile application usage
 4.4 Integrating FinTech comes with challenges
PwC · 2016 · pain_points
4.4 Integrating FinTech comes with challenges
 Surfing the FinTech wave – How to ensure your organisation is ready to take advantage of the FinTech people opportunity
PwC · 2016 · key_messages
Surfing the FinTech wave – How to ensure your organisation is ready to take advantage of the FinTech people opportunity
 Transfer Learning: From predicting everyday objects on ImageNet to detecting skin cancer
AirStreetCapital · case_study
Transfer Learning: From predicting everyday objects on ImageNet to detecting skin cancer
0.96 — AUC
 AI hardware as the new frontier
AirStreetCapital · industry_trends
AI hardware as the new frontier
18 Mins — Training time
 AI hardware as the new frontier
AirStreetCapital · industry_trends
AI hardware as the new frontier
25x — Training speedup
 AI hardware is especially helpful for deep learning
AirStreetCapital · industry_trends
AI hardware is especially helpful for deep learning
 AI hardware is especially helpful for deep learning
AirStreetCapital · other
AI hardware is especially helpful for deep learning
 AI hardware rate limits progress in today's deep learning era
AirStreetCapital · industry_trends
AI hardware rate limits progress in today's deep learning era
300,000x — Petaflop/s-day (Training)
 No wonder that GPUs have grown immensely in popularity amongst developers
AirStreetCapital · industry_trends
No wonder that GPUs have grown immensely in popularity amongst developers
+15X — Developer count / Downloads
 While GPUs are used extensively for training, they're not really needed for inference
AirStreetCapital · industry_trends
While GPUs are used extensively for training, they're not really needed for inference
 Processor clock frequencies are not getting faster and Moore’s Law can only take us so far
AirStreetCapital · industry_trends
Processor clock frequencies are not getting faster and Moore’s Law can only take us so far
 New architectures optimise for memory and compute to offer state-of-the-art performance running AI models
AirStreetCapital · industry_trends
New architectures optimise for memory and compute to offer state-of-the-art performance running AI models
242x — Inferences per second
 But GPUs and novel silicon are costly to rent per hour, which means progress is limited by financial resources
AirStreetCapital · industry_trends
But GPUs and novel silicon are costly to rent per hour, which means progress is limited by financial resources
$6.50 — Google Compute Engine price/hour
 Whilst more costly per hour, new silicon (e.g. Google's TPUv2) allows for faster model training at lower final costs (vs. NVIDIA V100)
AirStreetCapital · analyze_data
Whilst more costly per hour, new silicon (e.g. Google's TPUv2) allows for faster model training at lower final costs (vs. NVIDIA V100)
$55 — Cloud cost to reach 75.7% top-1 accuracy
 NVIDIA's datacenter business breaks $2B run-rate, is growing >100% year on year and accounts for almost 20% of their group revenue
AirStreetCapital · industry_trends
NVIDIA's datacenter business breaks $2B run-rate, is growing >100% year on year and accounts for almost 20% of their group revenue
$2B — Datacenter revenue
 NVIDIA's enterprise value has 10x in 3 years since the deep learning revolution ignited
AirStreetCapital · market_landscape
NVIDIA's enterprise value has 10x in 3 years since the deep learning revolution ignited
10x — Enterprise Value
 Intel's datacenter group accounts for 30% of the company's group revenue
AirStreetCapital · financial_analysis
Intel's datacenter group accounts for 30% of the company's group revenue
$19.1bn — Revenue
 Many companies are developing custom AI chips
AirStreetCapital · market_landscape
Many companies are developing custom AI chips
 Large technology companies are hedging their hardware suppliers, but there are few options to choose from
AirStreetCapital · competitive_analysis
Large technology companies are hedging their hardware suppliers, but there are few options to choose from
 Cloud giants are creating dedicated AI hardware and significantly growing their capex budgets
AirStreetCapital · industry_trends
Cloud giants are creating dedicated AI hardware and significantly growing their capex budgets
$76bn — Cloud capex ($mn)
 Traditional computer vision describes visual scenes by learning to detect objects ('nouns')
AirStreetCapital · context
Traditional computer vision describes visual scenes by learning to detect objects ('nouns')
 However, detecting objects in images is not enough to produce real scene understanding
AirStreetCapital · case_study
However, detecting objects in images is not enough to produce real scene understanding
 True scene understanding requires understanding actions ('verbs') and common sense
AirStreetCapital · case_study
True scene understanding requires understanding actions ('verbs') and common sense
 Building datasets for teaching machine learning models to understand video
AirStreetCapital · process_diagram
Building datasets for teaching machine learning models to understand video
 Deep learning models can actually understand the verbs as well as the nouns in video
AirStreetCapital · case_study
Deep learning models can actually understand the verbs as well as the nouns in video
 Machines can also understand visual scenes by learning to see from multiple viewpoints
AirStreetCapital · case_study
Machines can also understand visual scenes by learning to see from multiple viewpoints
 RL systems can learn goal-oriented behavior within simulated environments, i.e. games
AirStreetCapital · process_diagram
RL systems can learn goal-oriented behavior within simulated environments, i.e. games
 AlphaZero showed that a deep RL system can learn from scratch to beat Go champions
AirStreetCapital · case_study
AlphaZero showed that a deep RL system can learn from scratch to beat Go champions
 OpenAI's multi-agent RL system learns to play complex real-time strategy game, Dota2
AirStreetCapital · case_study
OpenAI's multi-agent RL system learns to play complex real-time strategy game, Dota2
 RL agents can also build their own world models and be trained within them
AirStreetCapital · case_study
RL agents can also build their own world models and be trained within them
906 — Average Score over 100 Random Tracks
 Fairness in machine learning: How do we ensure our models are not biased?
AirStreetCapital · industry_trends
Fairness in machine learning: How do we ensure our models are not biased?
 An example of biased machine learning systems: Stereotyping
AirStreetCapital · case_study
An example of biased machine learning systems: Stereotyping
 Another example of biased machine learning systems: Racial bias
AirStreetCapital · case_study
Another example of biased machine learning systems: Racial bias
 ML models have 5 types of allocation bias that stem from training data
AirStreetCapital · data_table
ML models have 5 types of allocation bias that stem from training data
 Like all software, ML models need to be debugged, but understanding them is hard
AirStreetCapital · other
Like all software, ML models need to be debugged, but understanding them is hard
 Explainability helps validate that ML models perform well for the “right” reasons
AirStreetCapital · case_study
Explainability helps validate that ML models perform well for the “right” reasons
 Next step: Justifying decisions in plain language and pointing to the evidence
AirStreetCapital · case_study
Next step: Justifying decisions in plain language and pointing to the evidence
 Understanding feature importance gives us high level insight into a model's behavior
AirStreetCapital · other
Understanding feature importance gives us high level insight into a model's behavior
 Imperceptible changes to data can alter a deep learning model's prediction
AirStreetCapital · other
Imperceptible changes to data can alter a deep learning model's prediction
99.3% — model confidence
 Imperceptible changes to data can alter a deep learning model's prediction
AirStreetCapital · case_study
Imperceptible changes to data can alter a deep learning model's prediction
 Adversarial attacks present serious safety challenges in the real world
AirStreetCapital · case_study
Adversarial attacks present serious safety challenges in the real world
 Improving deep learning model architecture requires iterative experimentation
AirStreetCapital · other
Improving deep learning model architecture requires iterative experimentation
 Improving deep learning model architecture requires iterative experimentation
AirStreetCapital · industry_trends
Improving deep learning model architecture requires iterative experimentation
80% — Top-1 accuracy
 AI to automate away AI engineers
AirStreetCapital · case_study
AI to automate away AI engineers
 Distributed “federated” learning to decentralise data acquisition and model training
AirStreetCapital · other
Distributed “federated” learning to decentralise data acquisition and model training
 Federated learning to decentralise data acquisition and model training
AirStreetCapital · process_diagram
Federated learning to decentralise data acquisition and model training
Next page →