Has Machine Learning AI hit the ceiling?

Forklift and Rack Storage


Artificial Intelligence (AI) has been an area of significant technological advancements in recent years. With the development of large machine learning (ML) models, AI has demonstrated unprecedented capabilities in various fields, including healthcare, finance, and education. However, the question arises: has AI hit the ceiling? The answer to this question may not be as straightforward as it seems.

In a recent interview, Open AI CEO, Sam Altman, highlighted the challenges of large ML models, stating that "the cost of training these models is growing exponentially, and we're starting to see diminishing returns." He also emphasized the importance of interpretability and the ethical implications of these models.

The current trend in AI is to create larger and more complex ML models. These models are designed to learn from vast datasets, making them more accurate and efficient. However, this approach has several limitations, primarily the need for significant computational power and the lack of interpretability. In addition, there is a growing concern about the ethical implications of such models, particularly regarding data privacy and bias.

The solution? Atomic Models

The idea of atomic models is not new. Atomic models are smaller, simpler models that are designed to perform a single task. These models are easier to interpret, more computationally efficient, and can be easily integrated into larger systems. However, the challenge is to create atomic models that are accurate enough to be used in complex systems.

The suggestion is not necessarily to replace large models entirely but to decouple them into smaller, more manageable atomic models. This approach can address several challenges posed by large models, such as interpretability, computational power, and ethical implications. Additionally, atomic models can be used to address specific tasks, making them more specialized and efficient.

Whereas the omnipresent ChatGPT is dealing with relatively simple textual data, and noting that this leads to enormous ML models, Sentispec's Logistics Productivity solutions deal with Visual Data which is infinitely more complex. As such I am sure you can imagine the resulting size of models using the same old methodology...

Hence, we have for a while been rapidly moving away from large, monolithic, models, towards networks of smaller, atomic, models which can be independently tested and validated against static "ground truth".

The nitty gritty of atomic models

One of the challenges of large models is their susceptibility to adversarial attacks. Adversarial attacks are intentional attacks on the ML model to manipulate its output. Atomic models, on the other hand, are more robust to such attacks due to their simplicity and specialized nature. By breaking down large models into atomic models, the overall system becomes more robust to adversarial attacks.

Another advantage of atomic models is that they can be trained on smaller datasets. Training large models on massive datasets can be time-consuming, costly, and may lead to overfitting. Atomic models, on the other hand, can be trained on smaller datasets, making the training process more efficient and cost-effective.

Furthermore, atomic models can be easily modified and updated, making them more adaptable to changing environments. In contrast, large models may require significant retraining and modifications to adapt to new tasks and datasets. Atomic models can also be integrated into larger systems, making them more scalable and versatile.

Decoupling large models into atomic models also addresses the issue of interpretability. Large models are often considered "black boxes," making it challenging to understand how they arrive at their decisions. Atomic models, on the other hand, are simpler and more transparent, making them easier to interpret and understand.

Lastly, breaking down large models into atomic models addresses ethical concerns regarding data privacy and bias. Large models often require massive amounts of data, which may not be readily available or may infringe on individuals' privacy. Atomic models, on the other hand, can be trained on smaller, more specific datasets, reducing the risk of data privacy violations.

Needless to say, Data Privacy and -Security is of paramount importance to Sentispec's Logistics Transparency solutions.

In conclusion

In conclusion, the question of whether AI has hit the ceiling is complex and multifaceted. However, the challenges posed by large ML models suggest that breaking them down into networks of atomic models may be necessary. Atomic models offer several advantages, such as interpretability, efficiency, robustness, and ethical considerations.

PS. There is nothing new here. As always in Computer Science, the solution to highly complex problems is to decompose them until you understand them.

Share this post

Stay Ahead in the Warehousing Game!
Never miss a beat. Subscribe to our blog and get the latest posts directly in your inbox as soon as we press “Publish”
By submitting this form, you agree to our privacy policy for the management of your data.

Recommended Articles

Forklift in Wharehouse
How to scale Predictable AI in logistics?

The use of AI in logistics can be cumbersome to scale because of the need for extensive training and huge data loads. But by prioritising only needed information and ignoring the rest, we can reduce a 10 MB sized problem into a 0.5-1 MB instead, meaning vastly reduced cost of training and operation as well […]

Security Cameras
The Fundamental Problem of Scaling Computer Vision

“To a man with a hammer, every problem looks like a nail.” The reason Computer Vision type applications and AI in general is so hard to scale is simple: The type of skills used to build the core of the solution are not the same as those required to maintain, release, deploy, and scale. In […]

Logistics Is Underserved by AI
Why Logistics Is Underserved by AI

Logistics isn’t something we tend to think about on an everyday basis. We order a thing on Amazon, and it arrives at our doorstep. We take our car to the shop, they change the oil and the windscreen viper. We don’t stop and think about how it got there, but indulge me for a moment […]

Ship Being Loaded with Containers
End-end Supply Chain Visibility

End-end supply chain visibility – the holy grail of logistics for any manufacturer or retailer. This is an extremely difficult problem to solve, but the potential benefits are immense. The reason it’s hard to achieve true end-end visibility is generally the number of partners involved in moving this around the globe through highly complicated transportation […]