Today, AI’s significant applications are being recognised in the world when it comes to solving complex problems. AI and its branch, deep learning have considerably contributed across the sector with machine translation, natural language processing and computer vision. And over the last few years, we have witnessed a spur in deep learning startups, but like any other software-based ones, these startups encounter some pitfalls, but these mistakes are a little unique to them.
Not Investing Enough in Data and Powerful Processors
The primary reasons for any deep learning startup to become successful is data and computation. GPUs bring down the time taken for computations, they reduce the weeks of computation time into a matter of hours and then there are TPUs which take even less time.
- Reducing the time taken to train the models is a significant mark when it comes to any company or startup, and that’s why it becomes imperative that one invests in GPUs smartly. One can buy their own GPUs or leverage Amazon Web Services or Google Cloud services.
- Another essential aspect when it comes to deep learning is the data. The odds are that the deep learning startup might not be successful at all if one doesn’t concentrate on data. There are some common mistakes that the startups need to avoid when it comes to data:
- Ignoring some of the unprocessed data completely and collecting unnecessary data can negatively impact models’ efficiency.
- Lack of Data: Perhaps, to avoid some mistakes, one can train the model on less data, but that creates problems too. One cannot train a model on 5-6 pictures of a Korean Jindo Dog breed to differentiate it in a pack of wolves.
- Lots of Data: But, training the model on too many dog pictures might eventually lead to the model identifying the fox as a dog breed. So, one needs to keep the balance between bias and variance.
A prime example is a bias in AI systems, a problem where even the blue-chip companies fail to provide clarity due to a lot of data that is being used; Amazon rekognition bias.
- Data quality: As much data one gives the system, one needs to keep in mind the quality of data too. Look at what happened with IBM’s Watson, with so much hype produced around it giving accurate advice on cancer treatment, at the end when it was reviewed, turn out it was giving erroneous advice.
Such mishaps when it comes to solving broad problems is a big risk and care must be taken about the quality of the data. The reason the output was erroneous was that they trained the model on a small number of hypothetical cancer patients and not the real ones.
Not Accounting for the Cloud Charges
Most AI applications look like regular software, but the centre of any AI application is a trained data model.
- The models perform many complex tasks like transcribe speech, generate natural language, and work similar to software as a service (SaaS). The bigger the software application and the more complex tasks they perform, the higher the bill they generate. So, one can presume how many bills AI might generate.
- When it comes to software startups, earlier, the cost of running the software on desktops or servers had to be paid by the buyer. In today’s SaaS dominant market, the software startups need to invest way less than they used to earlier because the cost has been pushed back to the vendor and most software companies pay AWS or Azure every month for the services they use.
- The AI system generates a lot of bills, even with simple training, retraining, and evaluating the model. Besides, the load as well as cost increases while using images, audio and video data types due to the need for high-performance computing.
- The cloud operations to scale the AI models globally aren’t the same across all the regions. So, as a result, the deep learning or machine learning startups always try to transfer the trained models across cloud regions which generate high ingress and egress costs to improve reliability and latency.
Expensive Data Cleansing
We all know that training the model once is not enough to achieve a state-of-the-art result, it has to be retrained for more accurate results.
- Training the model nowadays to achieve the state-of-the-art results involves a lot of manual cleaning and labelling of large datasets. And the process of manual cleaning and labelling is expensive and is one of the largest barriers the deep learning startups face.
Another area where human intervention is needed is where a lot of cognitive reasoning is required, the best example is the autonomous cars.
- Although as time passes, the AI systems are moving towards complete automation, which will significantly reduce the cost. However, these AI-based automation applications still need human intervention for years to come. Even if there is full automation achieved, it’s not clear how much the margin of cost and efficiency will improve, so this becomes a matter of whether one should invest towards processes like drift learning and active learning to enhance the ability.
- Not only expensive, the human intervention sometimes hinders the system’s creativity, but they might also do it by selecting what is essential for an algorithm to process or not using deep learning for a problem it can easily solve. Many times, deep learning is seen as overkill for many problems.
The costs incurred by human intervention and cloud are interdependent. Reducing one means an increase in another. Besides, startups aren’t the only ones who suffer from the problem of data cleaning, social media giants like Facebook still struggle to keep hate content and politically motivated content out of their platforms, even if they use state-of-the-art deep learning practices.
The Edge Cases
Many deep learning startups or AI startups suffer from edge cases. Users of AI-based apps or services can and will enter anything into an AI application thinking that it will take care of the rest. The users sometimes think that AI has super capabilities to process whatever data that has been put in; the ones who deal with the repercussions is the deep learning team of the company or the startup.
What happens in edge cases is that users end up giving huge and different amounts of input. Each customer input then generates data that is new to the system and something which it hasn’t seen before. Then the startups need to run a dedicated data collection and model fine-tuning for each of the customer engagement, which might reduce the edge cases but incurs a lot of costs until the model’s accuracy reaches a certain level. Situations like these ultimately hinder the AI system’s scalability.
Hiring the Right People
Tools like Keras, PyTorch and Tensorflow have made deep learning more accessible, and it has become relatively easy to get deep learning to work but, for something groundbreaking, more familiarity, high computation and in-depth knowledge is required.
One should take care that their deep learning team can build something beyond just simple copy-paste tasks. Rather than focusing on hiring Ph.Ds, they should strive to understand the personnel that is going to work for you.
Closing Thoughts
Although setting up a deep learning shop is a challenging endeavour, focusing and choosing the problem domain and trying to reduce the data complexity is recommended by some of the industry leaders. With the massive growth in the AI market and the expected increase from $9.5 billion in 2018 to $118.6 billion in 2025, to avoid some pitfalls, finding a niche problem and choosing not to work on broader problems like general text suggestions to avoid usual hindrance.
